Audience Level Data Scientists and Data Science Practitioners with basic to intermediate level experience in image recognition/object detection deep learning applications, basic Python programming skills.
Brief Description Deep Learning is revolutionizing both Data Science and Artificial Intelligence real-world applications. Yet, being the discipline so young, it’s not straightforward to understand both the reasoning at its core and its countless use cases. In this talk we will review the basics of Neural Networks, from the classical CNNs to the current state of the art, comparing them through real industry applications and highlighting pros and cons in a business setting.
Abstract / Summary Deep Learning has been on a hype roll for a few years. Being such a young discipline makes deep learning interesting, but also subject to misunderstandings. Every year, brand new architectures rise, taking over old ones and outperforming state of the art benchmarks for accuracy. Further, the applications of deep learning are at the core of some of the most advanced technologies like autonomous driving, personal assistants, and customer profiling. In such a context it is not straightforward to grasp what is at the core of deep learning itself, and what is common to all the architectures, neither to realize how concrete use cases can be tackled. Using Python, we’ll take the audience from the simplest neuron, the atom of the deep learning world, to the most recent architectures. We’ll achieve this using a simple Convolutional Neural Network as a building block and comparing that to the latest breakthroughs in image recognition. In doing this we’ll try to give an answer to the following questions: • Is deep learning actually useful in a business setting? • What about state of the art techniques in the Computer Vision field: are we just stacking more and more convolutional and pooling layers? We will then address real industry applications (e.g. the insurance sector) using analyzed techniques. This will include opening some of these so- called black box models and retraining them, at least partially, on our datasets, or building a complete brand new network from scratch, tailoring it according to the application needs and datasets characteristics.
in __on sabato 21 aprile at 17:45 **See schedule**